refactor(sae): move SST-2 SAE example into notebook tutorials

[phusroyal]

Summary 📝

Adds SAE feature visualization utilities and a notebook-based SST-2/GPT-2 SAE interpretation workflow. The PR makes it easier to inspect why one SAE feature behaves like positive or negative evidence by connecting three views: feature ranking, vocabulary-level logit effects, and token-level activations.

The SST-2 SAE example has also been moved out of examples/ and into the SAE notebook/tutorial flow, so the long-form interpretability walkthrough lives with the other notebook-based SAE material.

Details

• Adds plot_sae_feature_logit_effects(feature_id, decoder, unembedding, ...)

  • Projects one SAE decoder direction through the model unembedding.
  • Shows strongest negative and positive token effects.
  • Adds a signed vocabulary-wide logit-effect histogram.
  • Handles decoder/unembedding orientation explicitly.

• Adds plot_sae_token_activations(...)

  • Displays top token activations for one SAE feature.
  • Accepts either prepared activation rows or raw [N, seq] / [N, seq, features] arrays.
  • Supports BOS and padding masking.
  • Highlights readable token activations in compact rows.

• Replaces the old SST-2 SAE script with a notebook workflow:

  • moved from examples/sae_sst2_feature_enrichment.py
  • now lives as tutorials/sae/03_sst2_feature_enrichment.ipynb
  • loads SST-2, encodes GPT-2 SAE activations, ranks sentiment-selective features, and displays positive/negative feature visualizations.

• Updates SAE feature ranking:

  • readable-token filtering
  • smoothed label-token weights
  • sentence-level selectivity
  • contribution-ranked examples
  • compact output summaries

• Exports the new SAE plotting APIs from murano.plotting.

  • adds nltk to the notebook optional dependency check
  • regenerates uv.lock

Bugfixes 🐛

• SAE logit-effect plotting now validates decoder and unembedding shape/orientation instead of assuming one layout.
• Notebook target selection now fails loudly if either positive or negative candidates are missing.
• The notebook no longer stores huge printed rows or Plotly payloads by default, keeping the committed notebook lightweight.

Validation

• ruff format --check
• ruff check --fix --exit-non-zero-on-fix
• uv lock --check
• pyright src
• git diff --cached --check